mod
This commit is contained in:
l.gabrysiak
parent
3361683ac0
commit
d5049b651c
180
hft.py
180
hft.py
|
|
@ -11,12 +11,11 @@ import pytesseract
|
|||
from PIL import Image
|
||||
from collections import defaultdict
|
||||
from huggingface_hub import login
|
||||
from torch.utils.data import DataLoader
|
||||
|
||||
# Konfiguracja
|
||||
os.environ['TORCH_USE_CUDA_DSA'] = '1'
|
||||
os.environ["TOKENIZERS_PARALLELISM"] = "false"
|
||||
login(token="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") # Zastąp swoim tokenem
|
||||
login(token="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")
|
||||
|
||||
class SourceMapper:
|
||||
def __init__(self):
|
||||
|
|
@ -98,31 +97,31 @@ def prepare_dataset(directory, catalog_path, source_mapper):
|
|||
print(f"Rozpoznany typ dokumentu: {doc_type}")
|
||||
|
||||
if doc_type != "Opracowanie własne":
|
||||
# Ulepszone wyrażenie regularne dla różnych formatów
|
||||
articles = re.split(r'(?i)(Art[^\S\n]*\.?[^\S\n]*\d+[^\S\n]*\.?)', text)
|
||||
articles = re.split(r'(?i)(Art[\.\s]+\d+[\.\s]?)', text)
|
||||
articles = [a.strip() for a in articles if a.strip()]
|
||||
|
||||
print(f"Znaleziono {len(articles)//2} artykułów")
|
||||
print(f"Znaleziono {len(articles)} fragmentów")
|
||||
|
||||
# Generowanie większej liczby przykładów
|
||||
for i in range(0, len(articles)-1, 2):
|
||||
for chunk_size in [256, 512, 1024]: # Różne rozmiary chunków
|
||||
article_number = articles[i]
|
||||
article_content = articles[i+1]
|
||||
|
||||
if len(article_content) < 50:
|
||||
print(f"Pominięto zbyt krótki artykuł: {article_number}")
|
||||
continue
|
||||
chunks = [article_content[j:j+chunk_size] for j in range(0, len(article_content), chunk_size//2)]
|
||||
chunks = [c for c in chunks if len(c) > 100]
|
||||
|
||||
for chunk in chunks:
|
||||
source = f"{doc_type}, {article_number}"
|
||||
print(f"Dodano artykuł: {source}")
|
||||
|
||||
source_mapper.add_source(source)
|
||||
data.append({
|
||||
"text": f"{article_number} {article_content}",
|
||||
"text": f"{article_number} {chunk}",
|
||||
"source_idx": source_mapper.get_idx(source)
|
||||
})
|
||||
else:
|
||||
clean_text = re.sub(r'\s+', ' ', text).strip()
|
||||
chunks = [clean_text[i:i+512] for i in range(0, len(clean_text), 512)]
|
||||
for chunk_size in [256, 512, 768]: # Trzy różne rozmiary
|
||||
chunks = [clean_text[i:i+chunk_size] for i in range(0, len(clean_text), chunk_size//2)]
|
||||
chunks = [c for c in chunks if c.strip()]
|
||||
|
||||
for chunk in chunks:
|
||||
|
|
@ -130,7 +129,7 @@ def prepare_dataset(directory, catalog_path, source_mapper):
|
|||
"text": chunk,
|
||||
"source_idx": -1
|
||||
})
|
||||
print(f"Dodano {len(chunks)} chunków")
|
||||
print(f"Dodano {len(chunks)*3} chunków")
|
||||
|
||||
except Exception as e:
|
||||
print(f"Błąd podczas przetwarzania pliku: {str(e)}")
|
||||
|
|
@ -150,44 +149,32 @@ class CustomModel(nn.Module):
|
|||
def __init__(self, model_name, config):
|
||||
super().__init__()
|
||||
self.base_model = AutoModelForCausalLM.from_pretrained(model_name, config=config)
|
||||
self.source_embedding = nn.Embedding(1000, config.hidden_size, padding_idx=-1)
|
||||
self.source_embedding = nn.Embedding(10000, config.hidden_size, padding_idx=-1)
|
||||
|
||||
# Zamrożenie warstw bazowego modelu
|
||||
# Fine-tuning części modelu
|
||||
for param in self.base_model.parameters():
|
||||
param.requires_grad = False
|
||||
for param in self.base_model.get_output_embeddings().parameters():
|
||||
param.requires_grad = True
|
||||
for param in self.base_model.get_input_embeddings().parameters():
|
||||
param.requires_grad = True
|
||||
|
||||
def forward(self, input_ids=None, attention_mask=None, labels=None, source_idx=None, **kwargs):
|
||||
if source_idx is not None:
|
||||
valid_indices = torch.clamp(source_idx, 0, self.source_embedding.num_embeddings-1)
|
||||
|
||||
source_embeds = torch.nn.functional.normalize(
|
||||
self.source_embedding(valid_indices),
|
||||
p=2,
|
||||
dim=-1
|
||||
).unsqueeze(1)
|
||||
|
||||
source_embeds = self.source_embedding(valid_indices).unsqueeze(1)
|
||||
inputs_embeds = self.base_model.get_input_embeddings()(input_ids) + source_embeds
|
||||
|
||||
return self.base_model(
|
||||
inputs_embeds=inputs_embeds,
|
||||
attention_mask=attention_mask,
|
||||
labels=labels,
|
||||
**kwargs
|
||||
)
|
||||
|
||||
return self.base_model(
|
||||
input_ids=input_ids,
|
||||
attention_mask=attention_mask,
|
||||
labels=labels,
|
||||
**kwargs
|
||||
)
|
||||
return self.base_model(inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels, **kwargs)
|
||||
return self.base_model(input_ids=input_ids, attention_mask=attention_mask, labels=labels, **kwargs)
|
||||
|
||||
def generate(self, *args, **kwargs):
|
||||
return self.base_model.generate(*args, **kwargs)
|
||||
|
||||
class CustomTrainer(Trainer):
|
||||
def __init__(self, *args, **kwargs):
|
||||
self.tokenizer = kwargs.pop('tokenizer', None)
|
||||
super().__init__(*args, **kwargs)
|
||||
|
||||
def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
|
||||
labels = inputs.pop("labels")
|
||||
source_idx = inputs.pop("source_idx", None)
|
||||
|
|
@ -195,72 +182,54 @@ class CustomTrainer(Trainer):
|
|||
return (outputs.loss, outputs) if return_outputs else outputs.loss
|
||||
|
||||
def evaluate(self):
|
||||
val_questions = {
|
||||
"art1": "Jakie są prawa pracownika według art. 1?",
|
||||
"art2": "Kto jest pracownikiem według art. 2?",
|
||||
"art3": "Jakie są obowiązki pracodawcy według art. 3?"
|
||||
}
|
||||
questions = [
|
||||
"Jakie są prawa pracownika według art. 1?",
|
||||
"Kto jest pracownikiem według art. 2?",
|
||||
"Jakie są obowiązki pracodawcy według art. 3?"
|
||||
]
|
||||
|
||||
model.eval()
|
||||
results = {}
|
||||
|
||||
for key, question in val_questions.items():
|
||||
result = self.generate_answer(question)
|
||||
results[key] = result
|
||||
|
||||
print("\nWyniki walidacji:")
|
||||
for key, val in results.items():
|
||||
print(f"\n{val_questions[key]}")
|
||||
print(f"Odpowiedź: {val['answer'][:200]}...")
|
||||
print(f"Źródła: {val['sources']}")
|
||||
print("\n" + "="*50 + "\nEWALUACJA\n" + "="*50)
|
||||
for q in questions:
|
||||
result = self.generate_answer(q)
|
||||
print(f"\nPYTANIE: {q}")
|
||||
print(f"ODPOWIEDŹ: {result['answer'][:500]}")
|
||||
print(f"ŹRÓDŁA: {', '.join(result['sources'])}")
|
||||
print("-"*80)
|
||||
|
||||
return {"loss": 0.0}
|
||||
|
||||
def generate_answer(self, question):
|
||||
tokenizer = self.tokenizer
|
||||
model = self.model
|
||||
device = model.base_model.device
|
||||
|
||||
prompt = f"[PYTANIE PRAWNE] {question} [KONTEKST]"
|
||||
|
||||
inputs = tokenizer(
|
||||
prompt,
|
||||
inputs = self.tokenizer(
|
||||
f"[PYTANIE] {question} [KONTEKST]",
|
||||
return_tensors="pt",
|
||||
truncation=True,
|
||||
max_length=512
|
||||
).to(device)
|
||||
).to(self.model.base_model.device)
|
||||
|
||||
with torch.no_grad():
|
||||
outputs = model.generate(
|
||||
outputs = self.model.generate(
|
||||
**inputs,
|
||||
max_new_tokens=150,
|
||||
temperature=0.3,
|
||||
top_k=50,
|
||||
top_p=0.95,
|
||||
repetition_penalty=1.8,
|
||||
max_new_tokens=200,
|
||||
temperature=0.5,
|
||||
top_p=0.9,
|
||||
repetition_penalty=2.0,
|
||||
num_beams=3,
|
||||
no_repeat_ngram_size=4,
|
||||
early_stopping=True,
|
||||
pad_token_id=tokenizer.eos_token_id
|
||||
no_repeat_ngram_size=3
|
||||
)
|
||||
|
||||
answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
||||
answer = answer.replace(prompt, "").strip()
|
||||
answer = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
|
||||
answer = answer.split("[KONTEKST]")[-1].strip()
|
||||
|
||||
sources = set()
|
||||
for match in re.finditer(r'(?i)art\.?\s*\d+\.?', answer):
|
||||
article_ref = match.group(0).strip().rstrip('.')
|
||||
for source in self.model.source_mapper.idx_to_source.values():
|
||||
for match in re.finditer(r'(?i)art\.?\s*\d+', answer):
|
||||
article_ref = match.group(0).strip()
|
||||
for idx, source in self.model.source_mapper.idx_to_source.items():
|
||||
if article_ref.lower() in source.lower():
|
||||
sources.add(source)
|
||||
|
||||
return {
|
||||
"answer": answer,
|
||||
"sources": list(sources) if sources else ["Opracowanie własne"]
|
||||
}
|
||||
return {"answer": answer, "sources": list(sources)}
|
||||
|
||||
def main():
|
||||
# Inicjalizacja
|
||||
source_mapper = SourceMapper()
|
||||
model_name = "crumb/nano-mistral"
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||||
|
|
@ -271,70 +240,53 @@ def main():
|
|||
data = prepare_dataset("files", catalog_path, source_mapper)
|
||||
|
||||
if not data:
|
||||
print("\nBrak danych do treningu! Sprawdź pliki w katalogu 'files' i diagnostykę powyżej.")
|
||||
print("\nBrak danych do treningu!")
|
||||
return
|
||||
|
||||
dataset = Dataset.from_list(data)
|
||||
|
||||
def tokenize_function(examples):
|
||||
tokenized = tokenizer(
|
||||
def tokenize(examples):
|
||||
return tokenizer(
|
||||
examples["text"],
|
||||
truncation=True,
|
||||
padding="max_length",
|
||||
max_length=512,
|
||||
return_tensors="pt"
|
||||
)
|
||||
return {
|
||||
"input_ids": tokenized["input_ids"][0],
|
||||
"attention_mask": tokenized["attention_mask"][0],
|
||||
"labels": tokenized["input_ids"][0].clone(),
|
||||
"source_idx": examples["source_idx"]
|
||||
}
|
||||
|
||||
tokenized_dataset = dataset.map(tokenize_function, batched=False)
|
||||
tokenized_dataset = dataset.map(tokenize, batched=True, batch_size=16)
|
||||
|
||||
def custom_collate_fn(features):
|
||||
return {
|
||||
"input_ids": torch.stack([torch.tensor(f["input_ids"]) for f in features]),
|
||||
"attention_mask": torch.stack([torch.tensor(f["attention_mask"]) for f in features]),
|
||||
"labels": torch.stack([torch.tensor(f["labels"]) for f in features]),
|
||||
"source_idx": torch.tensor([f["source_idx"] for f in features], dtype=torch.long)
|
||||
}
|
||||
|
||||
# Model
|
||||
config = AutoModelForCausalLM.from_pretrained(model_name).config
|
||||
model = CustomModel(model_name, config)
|
||||
model.source_mapper = source_mapper # Dodanie mapowania źródeł do modelu
|
||||
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
||||
model.to(device)
|
||||
|
||||
# Trening
|
||||
training_args = TrainingArguments(
|
||||
output_dir="./results",
|
||||
num_train_epochs=5,
|
||||
num_train_epochs=8,
|
||||
per_device_train_batch_size=4,
|
||||
gradient_accumulation_steps=2,
|
||||
learning_rate=1e-5,
|
||||
gradient_accumulation_steps=8,
|
||||
learning_rate=5e-6,
|
||||
weight_decay=0.01,
|
||||
warmup_ratio=0.1,
|
||||
fp16=torch.cuda.is_available(),
|
||||
logging_steps=10,
|
||||
logging_steps=50,
|
||||
save_strategy="epoch",
|
||||
evaluation_strategy="steps",
|
||||
eval_steps=500,
|
||||
eval_strategy="no",
|
||||
report_to="none",
|
||||
remove_unused_columns=False
|
||||
)
|
||||
|
||||
model = CustomModel(model_name, AutoModelForCausalLM.from_pretrained(model_name).config)
|
||||
model.source_mapper = source_mapper
|
||||
model.to("cuda" if torch.cuda.is_available() else "cpu")
|
||||
|
||||
trainer = CustomTrainer(
|
||||
model=model,
|
||||
args=training_args,
|
||||
train_dataset=tokenized_dataset,
|
||||
data_collator=custom_collate_fn,
|
||||
tokenizer=tokenizer
|
||||
)
|
||||
|
||||
print("\nRozpoczęcie treningu...")
|
||||
trainer.train()
|
||||
|
||||
print("\nKońcowa ewaluacja...")
|
||||
trainer.evaluate()
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
|
|
|||
Loading…
Reference in New Issue